3D cone traces #2641

etpinard · 2018-05-16T18:48:00Z

Cones are coming to plotly.js

Things we need to decide on

How to name this trace? I chose cone, like a cone trace, similar to surface. But since a cone trace has often many cones, maybe cones would be better. cone3d similar to mesh3d could work, but I think the 3d is not necessary, quiver is more popular in 2d. Decision: we're going with cone.
How to name the meshgrid attribute? I chose vx, vy and vz (v for vector), but maybe meshgrid.x, meshgrid.y and meshgrid.z would be best. Note that these attributes are optional. Decision: ~~going with cones.(x|y|z) in cb7ef43 which will allows us to configure non-cartesian-product meshgrids if we want down the road.~~ Removed from this 1st implementation -> 5a42de0
Should the colorbar show normalize vector norm (i.e. 0 to 1) values or absolute vector norms? Maybe we need an attribute to enable both behaviors? Decision: going with absolut vector norms in a2db567
The color values (i.e. intensity) for the cones represent the vector norm which is positive definite. Maybe we should restrict cmin and cmax to be positive also? Moreover, maybe we should make Viridis the default colorscale. Decision: Let's not restrict cmin, cmax in case we add an intensity field to color the cones by something other than their vector norm down the road. No go for Viridis default until v2.

Things we need to fix/add

Computed trace bounds are off. They only consider the cones' tip, ignoring their lengths and directions. This can make autoranged axes look weird:

Update: Ok, but not great solution done in db8222b and improved multiple times in subsequent commits.

Need to plug in size settings (attributes sizeref and sizemode) - done in 863b0da
Hover doesn't work - done in 71e46af
Image generation fails in imagetest (works fine in orca) - Update: non-perfect solution in aaf7249 where gl3d_cone-* mocks are tests using in orca during ./tasks/noci_test.sh.

Things that work ok, but could improve

We currently call cone2mesh twice which can be slow (~300ms for 20k cones) per .plot call. Once in calc to get the colors (i.e. intensity) values and another in convert to compute the correctly scaled coordinate (note the the scene's scale isn't available during calc). I would be nice to split gl-cone3d's createConePlot into two subroutines. Update: not any more after a2db567

Things that could wait future iterations

make x, y and z optional, generate linspaces when omitted
plug in gl-mesh3d's lighting settings - done in 3dd6cf5
~~[ ] providing an underdetermined meshgrid can lead to exceptions in gl-cone3d's createConePlot~~ not an issue until we figure how to implement our version of meshgrid.

- unfortunately they don't render correctly in `imagetest`

etpinard · 2018-05-16T18:49:56Z

forgot to cc @alexcjohnson @jackparmer and @kig who maybe could help out ✅ a few checkboxes.

alexcjohnson · 2018-05-16T19:52:33Z

Cool!

How to name this trace? I chose cone, like a cone trace, similar to surface. But since a cone trace has often many cones, maybe cones would be better. cone3d similar to mesh3d could work, but I think the 3d is not necessary, quiver is more popular in 2d.

Agreed - a cone is a 3D shape, it would have no meaning in 2D, so just cone - 3d is unnecessary. We don't pluralize any of our other types (parcoords looks plural, but it's "coordinates" that's plural, not the lines connecting these coordinates, which are really the objects on the plot)

How to name the meshgrid attribute? I chose vx, vy and vz (v for vector), but maybe meshgrid.x, meshgrid.y and meshgrid.z would be best. Note that these attributes are optional.

I think we should have the required data be the vector field (x, y, z) -> (u, v, w), then probably we need a few ways to specify whether and how to interpolate cones onto that field (and I think this is what you're already doing, but in case it's not clear already all of these arrays are assumed to be 1D - 3D leads to too many ambiguities as well as extra headaches for our back end):

If you add nothing else, draw cones at exactly the specified positions (x, y, z)
You can provide 3 more data arrays, then the cones are interpolated onto these positions. Perhaps like MATLAB we call these (cx, cy, cz)? or cones: {x, y, z}?
It would also be nice to be able to specify just a grid for each dimension; what should this look like? I don't think it can just be x0, dx etc because we need both bounds; x0, dx, x1 would do this but we have no precedent for that. For histograms and contours we have {start, end, size}, so how about cones: {xgrid: {start, end, size}, ygrid: {start, end, size}, zgrid: {start, end, size}}? start and end could be optional, then we would just bound the grid to the data.
In principle it seems like you could mix and match these per dimension... like you provide cones: {x, y, zgrid} and the result is cones explicitly positioned in x and y, then repeated for every value implied by zgrid. But that could get confusing, I could imagine people wanting to think of cones.x, cones.y, and cones.z each as an independent array, ie cone positions are the cartesian product of the three rather than element-by-element combinations. Too many possibilities, if you want to start with just the first option that's fine, we can add the others later if there's demand.

Should the colorbar show normalize vector norm i.e. 0 to 1) values or absolute vector norms? Maybe we need an attribute to enable both behaviors?

Absolute vector norm seems far more useful. Leave this attribute out and let people normalize on their own, until someone complains. The hover data is always going to show the absolute vector anyway, right? Would be weird if this disagreed with the colorscale.

The color values (i.e. intensity) for the cones represent the vector norm which is positive definite. Maybe we should restrict cmin and cmax to be positive also? Moreover, maybe we should make Viridis the default colorscale.

Much as I like Viridis, lets not diverge from the default colorscales we have for all other traces until we can change them as well.

Is there a use case to supply color as an independent data array? Perhaps someone wants to color just by w (vertical flux) or ∇∙v or ∇×v or who knows what... If we allow that, cmin and cmax could be anything. Anyway I don't see a need to explicitly force them to be positive even when color is the norm, they'd be positive by default and only go negative if the user explicitly chooses that.

alexcjohnson · 2018-05-16T19:59:39Z

Computed trace bounds are off. They only consider the cones' tip, ignoring their lengths and directions.

Are the cones positioned with their tips at the data positions? That seems odd, I'd have thought the data position would go at its center of mass, ie 3/4 of the way from the tip to the base.

etpinard · 2018-05-16T20:01:11Z

Are the cones positioned with their tips at the data positions?

Yes. That's how it works currently.

alexcjohnson · 2018-05-16T20:03:32Z

Huh, MATLAB places the cones with their tails at the data positions. That also seems wrong to me...

alexcjohnson · 2018-05-16T20:20:30Z

src/traces/cone/attributes.js

+            'If *scaled*, `sizeref` scales such that the reference cone size',
+            'for the maximum vector magnitude is 1.',
+            'If *absolute*, `sizeref` scales such that the reference cone size',
+            'for vector magnitude 1 is one grid unit.'


What do you mean by a grid unit here? Is it one unit of the (x, y, z) positions or the spacing between adjacent cones?

Almost always the vectors have different units from the position space they're displayed in - so I'd think if you specify nothing at all about the cone sizing, they should end up scaled so the largest one is about as big as the smallest spacing between adjacent cones.

There's definitely a use case for 'absolute' - the user has pre-calculated everything and they want each cone to go from one precise (x, y, z) to another precise (x, y, z). I'm not sure there's a use case though for scaling to one unit in position, I'd think 'scaled' should refer to the minimum adjacent cone distance.

To be honest, I'm not sure. I copied this from https://github.com/gl-vis/gl-cone3d#constructor I'll have to inspect

https://github.com/gl-vis/gl-cone3d/blob/c675b25e5991e63d0ad23ed24d712c496c28cd72/cone.js#L138-L246

a little bit more.

I'll have a new look at the cone sizing parameters. Pre-meshgrid they were based on the position space units and scaled so that two cones with maximum norm, pointing to the same direction, would have their tip and base touch. Like <|<|. That would be with coneSize 2. With coneSize 1, you'd get no overlapping cones even when they're |><|.

The absolute sizing ditched the automatic cone scaling and used the cone vector norm to figure out the size of the cone model, so that a norm of 1 with absolute size factor 1 results in a cone that's one unit tall in the position unit space. Absolute size factor 2 => 2 unit tall cone, etc.

With the meshgrid sampling (and the realization that the cone positions may not be on a nice regular grid), I changed the automatic cone sizing to scale based on the smallest distance between two adjacent cone positions (adjacent in the sense that they've got successive indices in the positions array). This is not the right thing to do though, doing a smallest distance between any two cone positions would be, but that's O(n^2) with a naive algo.

Another idea floated on the sizing was to have pixel-based sizing as well, which'd probably work like "invert the projection, view and model transform matrices for cone model, scale cone model according to viewport size" but would result in ortho camera cones which might be confusing. (And it sounds like a bit of a pain to implement.)

@kig that all sounds great. I guess we only need to worry about the smallest-distance calculation when using the (x, y, z) data to position the cones; With the cartesian product of x/y/z interpolation grids each dimension should be pretty compact, even if the values are out of order. But I guarantee at some point we'll see a data set where the closest points are not successive. I don't know if there's a JS implementation already available but it's possible to do it in O(n log(n)^2) http://www.cs.ucsb.edu/~suri/cs235/ClosestPair.pdf

Lets not worry about pixel-based sizing, at least not for now. I guess I can see using this if we also do automatic grid spacing, ie when you zoom in and the cone spacing gets larger, at some point we decrease the grid size so we draw more cones between the existing ones. That's definitely not a v1 feature 😅

alexcjohnson · 2018-05-16T20:36:29Z

We currently call cone2mesh twice which can be slow (~300ms for 20k cones) per .plot call. Once in calc to get the colors (i.e. intensity) values and another in convert to compute the correctly scaled coordinate

If it's just for color (which, per #2641 (comment), might not not even come from the cone lengths), can't we just find the bounds of the data provided, as a simple loop over the cone data? That should be a heck of a lot faster (quick test: ~5ms for 20k data points) and, I might argue, more faithful to the data than basing the color bounds on the interpolated data used in the actual cones.

... using an invisible gl-scatter3d trace

etpinard · 2018-05-16T21:34:56Z

src/traces/cone/convert.js

+    var meshData = convert(scene, data);
+    var mesh = createConeMesh(gl, meshData);
+
+    var pts = createScatterPlot({


a little bit hacky and not very efficient, but this works:

Nice, that was fast! Is it possible to get (u, v, w) in the hovertext as well?

It's possible!

but maybe someone would argue that 6 fields in one hover label in too much by default. Maybe we could just show the vector norm value? Or maybe have hoverinfo: 'x+y+z+norm' by default and optionally up to: x+y+z+u+v+w+norm+text

'x+y+z+norm' by default and optionally up to: x+y+z+u+v+w+norm+text

I like it!

done in e4a2035

... w/o using cone2mesh, and use that the find cmin & cmax which now correspond to "absolute" vector norm. This make cone/helpers.js obsolete.

... during autorange / scene bounds computation. - This pad values is a "worse case" solution (i.e. it pads too much for most situations), but gives nice visual results (I think).

etpinard · 2018-05-17T03:24:50Z

src/traces/cone/calc.js


-    colorscaleCalc(trace, meshData.vertexIntensity, '', 'c');
+    colorscaleCalc(trace, [normMin, normMax], '', 'c');


Shoutout to @alexcjohnson for the tip!

etpinard · 2018-05-17T03:28:38Z

src/plots/gl3d/scene.js

+                var obj = objects[j];
+                var objBounds = obj.bounds;
+                // add worse-case pad for cone traces
+                var pad = (obj._trace.data._normMax || 0) * dataScale[i];


This is what I came up with to solve the "cones out of scene" problem.

Plotly.newPlot(gd, [{ type: 'cone', "x": [1, 2, 3], "y": [1, 2, 3], "z": [1, 2, 3], "u": [1, 0, 0], "v": [0, 3, 0], "w": [0, 0, 2], }])

now gives:

Looks good! I suppose based on gl-vis/gl-cone3d#6 we could drop it to 3/4 of that value, though then if we make that configurable it would depend on the offset... more trouble than it's worth, let's keep what you have here.

- x/y/z always refers vector positions (their length should match u/v/w) - if nothing else provided, x/y/z correspond to the cone position also, - to set cone position, use `cones.(x|y|z)` for now - update attribute descriptions and trace meta info

etpinard · 2018-05-17T23:07:37Z

I still need to fixup this case brought up by @alexcjohnson

// seems like something’s off with the axis padding - cones at 1 & 3 give axis ranges `[-2.46, 6.46]`? Shouldn’t that be `[0, 4]` if we’re padding by one full cell?

Plotly.newPlot(gd,[{type: 'cone', x: [1, 1, 1, 1, 3, 3, 3, 3], y: [1, 1, 3, 3, 1, 1, 3, 3], z: [1, 3, 1, 3, 1, 3, 1, 3],
u: [1, 2, 3, 4, 1, 2, 3, 4], v: [1, 1, 2, 2, 3, 3, 4, 4], w: [4, 4, 1, 1, 2, 2, 3, 3]}])

but other than, this PR seems ready. Dropping the discussion needed, tagging as reviewable.

etpinard · 2018-05-18T01:14:13Z

src/plots/gl3d/scene.js

@@ -508,7 +508,7 @@ proto.plot = function(sceneData, fullLayout, layout) {
                var obj = objects[j];
                var objBounds = obj.bounds;
                // add worse-case pad for cone traces
-                var pad = (obj._trace.data._normMax || 0) * dataScale[i];
+                var pad = 0.75 * (obj._trace.data._compMax || 0) * dataScale[i];


This gives a better estimate. The data off #2641 (comment) now gives:

Great! Yeah, I was wrong with [0, 4] - this looks awesome.

hmm, this is still not quite right though. If I double all the u, v, w values, the cones don't change (as they shouldn't) but the padding increases.

var t={type: 'cone', x: [1, 1, 1, 1, 3, 3, 3, 3], y: [1, 1, 3, 3, 1, 1, 3, 3], z: [1, 3, 1, 3, 1, 3, 1, 3], u: [1, 2, 3, 4, 1, 2, 3, 4], v: [1, 1, 2, 2, 3, 3, 4, 4], w: [4, 4, 1, 1, 2, 2, 3, 3]} // the resulting image should be independent of what I multiply by here function m(v) { return v*2; } t.u = t.u.map(m); t.v = t.v.map(m); t.w=t.w.map(m); Plotly.newPlot(gd,[t], {scene: {camera: {eye: {x: -0.01, y: 2.52, z: 0.42}}}, width: 800, height: 800})

v * 1 (as above):

v * 2:

I don't need to tell you what happens with v / 10 do I? 😉

latest attempt: 3d26d47

alexcjohnson · 2018-05-18T13:01:38Z

test/jasmine/tests/cone_test.js

+            delete gd.data[0][k];
+
+            supplyAllDefaults(gd);
+            expect(gd._fullData[0].visible).toBe(!k, 'missing array ' + k);


Clever, but maybe a little too compact to include '' in this set 😈

ok, fixed in 8dca9ae

etpinard · 2018-05-18T16:50:06Z

test/jasmine/tests/cone_test.js

+        })
+        .then(function() {
+            _assertAxisRanges('after adding one cone outside range but with norm-0',
+                [-0.45, 6.45], [-0.45, 6.45], [-0.45, 6.45]


@alexcjohnson unfortunately I can't get this to "stick" to the previous range[0] values without breaking other gl3d mocks that depend on

plotly.js/src/plots/gl3d/scene.js

Lines 479 to 483 in fb75158

for(j = 0; j < objects.length; j++) {

var objBounds = objects[j].bounds;

sceneBounds[0][i] = Math.min(sceneBounds[0][i], objBounds[0][i] / dataScale[i]);

sceneBounds[1][i] = Math.max(sceneBounds[1][i], objBounds[1][i] / dataScale[i]);

}

which scales the autorange bounds with the dataScale:

plotly.js/src/plots/gl3d/scene.js

Lines 394 to 407 in fb75158

var dataScale = [1, 1, 1];

for(j = 0; j < 3; ++j) {

if(dataBounds[0][j] > dataBounds[1][j]) {

dataScale[j] = 1.0;

}

else {

if(dataBounds[1][j] === dataBounds[0][j]) {

dataScale[j] = 1.0;

}

else {

dataScale[j] = 1.0 / (dataBounds[1][j] - dataBounds[0][j]);

}

}

}

Hopefully this iteration is close enough.

As a side note, we should spend some time replacing this dataScale hack. This thing was added to make date axes work properly in 3D. Timestamps don't fit in a Float32Array, so we can't just convert the datestrings to timestamps, but we do this better in scattergl, where we use Float64Array or split the data into two Float32Arrays when not supported.

New issue about the dataScale mess -> #2646

alexcjohnson · 2018-05-18T20:26:05Z

src/traces/cone/convert.js

@@ -107,6 +108,9 @@ function convert(scene, trace) {
    meshData.fresnel = trace.lighting.fresnel;
    meshData.opacity = trace.opacity;

+    // stash autorange pad value
+    trace._pad = anchor2coneSpan[trace.anchor] * meshData.vectorScale * trace._normMax * trace.sizeref;


Awesome, I think we have a winner now! I don't see any cases (As I change the cone spacing, cone count in each direction, and vector length) where this fails to encompass all the cones, nor any where the padding looks unpleasantly large. 🏆

alexcjohnson · 2018-05-18T20:28:02Z

test/jasmine/tests/cone_test.js

@@ -102,7 +102,7 @@ describe('@gl Test cone autorange:', function() {
        })
        .then(function() {
            _assertAxisRanges('scaled up',
-                [-0.39, 4.39], [-0.39, 4.39], [-0.39, 4.39]
+                [-0.66, 4.66], [-0.66, 4.66], [-0.66, 4.66]


Can you add one more test, to 🔒 this in a little more since it went through so many iterations: scale up the x/y/z arrays. At least one of the previous iterations had a problem with this but otherwise looked pretty good.

there's this one:

plotly.js/test/jasmine/tests/cone_test.js

Lines 121 to 144 in abb11e0

var trace = fig.data[0];

var x = trace.x.slice();

x.push(5);

var y = trace.y.slice();

y.push(5);

var z = trace.z.slice();

z.push(5);

var u = trace.u.slice();

u.push(0);

var v = trace.v.slice();

v.push(0);

var w = trace.w.slice();

w.push(0);

return Plotly.restyle(gd, {

x: [x], y: [y], z: [z],

u: [u], v: [v], w: [w]

});

})

.then(function() {

_assertAxisRanges('after adding one cone outside range but with norm-0',

[-0.72, 6.72], [-0.72, 6.72], [-0.72, 6.72]

);

am I missing anything?

It's close, but there's a difference between adding another point that doesn't change the minimum spacing between cones, and expanding the whole x/y/z space and therefore expanding the minimum spacing as well.

got it. Thanks!

done in -> 2c08357

alexcjohnson · 2018-05-18T20:53:43Z

Alright! I'm happy if you are! 💃 💃 💃

etpinard · 2018-05-18T21:06:43Z

Ok. Merging this thing!

Shoutout to @kig for this feature 🎉

Thanks very much @alexcjohnson for the help making this plotly-proof! 💪

This will be part of v1.38.0 set to be released on Tuesday. 🚀

by the way @kig I published gl-cone3d@1.0.0 to npm.

etpinard added 4 commits May 16, 2018 12:37

scaffold cone

ec64895

move 3d parseColorScale to gl_format_color

5736290

get cone colorbar and vx/vy/vz mesh grid to work

cb7cb7f

add three cone mocks

89eaf11

- unfortunately they don't render correctly in `imagetest`

etpinard added type: feature status: has TODOs status: discussion needed labels May 16, 2018

etpinard added this to the v1.38.0 milestone May 16, 2018

jackparmer mentioned this pull request May 16, 2018

3D Support for Streamline/Quiver Plots plotly/plotly.py#861

Closed

implement 'sizemode' and 'sizeref'

863b0da

add lib/cone.js + add cone to gl3d bundle

fa32a74

alexcjohnson reviewed May 16, 2018

View reviewed changes

hook in hover labels for cones

71e46af

... using an invisible gl-scatter3d trace

etpinard commented May 16, 2018

View reviewed changes

etpinard added 3 commits May 16, 2018 23:18

clean up scene computeTraceBounds

a2c3694

find min/max u/v/w norm in calc directly

a2db567

... w/o using cone2mesh, and use that the find cmin & cmax which now correspond to "absolute" vector norm. This make cone/helpers.js obsolete.

add "pad" using max u/v/w norm around cone bounds

db8222b

... during autorange / scene bounds computation. - This pad values is a "worse case" solution (i.e. it pads too much for most situations), but gives nice visual results (I think).

etpinard commented May 17, 2018

View reviewed changes

etpinard mentioned this pull request May 17, 2018

Cones are stretched out when axis spans don't match gl-vis/gl-cone3d#4

Closed

etpinard added 4 commits May 17, 2018 10:15

add 'anchor' to mimic gl-cone3d's coneOffset

55700b8

add cone-specific hoverinfo flags

e4a2035

fixup cmin/cmax -> vertexIntensityBounds

a02dd11

etpinard added 3 commits May 17, 2018 17:58

add some cone tests

d5d6f33

add hover label test for cones

b7465fb

comment out cones.(x|y|z) attributes for now

5a42de0

etpinard added status: reviewable and removed status: discussion needed status: has TODOs labels May 17, 2018

better pad value for autoranged cones

04d3d91

etpinard commented May 18, 2018

View reviewed changes

revert to current gl-cone3d master

2a45dae

alexcjohnson reviewed May 18, 2018

View reviewed changes

etpinard added 4 commits May 18, 2018 10:27

factor out visible: false converse test

8dca9ae

fixup cone autorange pad computation

3d26d47

fixup pad autorange (again)

5a59bc7

pass gl-mesh3d lighting params

3dd6cf5

etpinard commented May 18, 2018

View reviewed changes

etpinard added 2 commits May 18, 2018 16:02

use gl-cone2d "vectorScale" to get better autorange pad value

fb2ec1e

scaled pad value by normMax only under sizemode: 'scaled'

74ecbf8

alexcjohnson reviewed May 18, 2018

View reviewed changes

sub in npm version for gl-cone3d

abb11e0

alexcjohnson reviewed May 18, 2018

View reviewed changes

etpinard added 2 commits May 18, 2018 16:49

fixup tests post 74ecbf8

8e8f5d6

add autorange test case for "scaled up the x/y/z arrays"

2c08357

etpinard mentioned this pull request May 18, 2018

Replace "dataScale" hack in 3D subplots #2646

Closed

etpinard merged commit df67a27 into master May 18, 2018

etpinard deleted the cone-traces branch May 18, 2018 21:07

This was referenced May 21, 2018

Fix single cone #2647

Merged

3D streamtube traces #2658

Merged


		colorscaleCalc(trace, meshData.vertexIntensity, '', 'c');
		colorscaleCalc(trace, [normMin, normMax], '', 'c');

	for(j = 0; j < objects.length; j++) {
	var objBounds = objects[j].bounds;
	sceneBounds[0][i] = Math.min(sceneBounds[0][i], objBounds[0][i] / dataScale[i]);
	sceneBounds[1][i] = Math.max(sceneBounds[1][i], objBounds[1][i] / dataScale[i]);
	}

	var dataScale = [1, 1, 1];
	for(j = 0; j < 3; ++j) {
	if(dataBounds[0][j] > dataBounds[1][j]) {
	dataScale[j] = 1.0;
	}
	else {
	if(dataBounds[1][j] === dataBounds[0][j]) {
	dataScale[j] = 1.0;
	}
	else {
	dataScale[j] = 1.0 / (dataBounds[1][j] - dataBounds[0][j]);
	}
	}
	}

	var trace = fig.data[0];

	var x = trace.x.slice();
	x.push(5);
	var y = trace.y.slice();
	y.push(5);
	var z = trace.z.slice();
	z.push(5);
	var u = trace.u.slice();
	u.push(0);
	var v = trace.v.slice();
	v.push(0);
	var w = trace.w.slice();
	w.push(0);

	return Plotly.restyle(gd, {
	x: [x], y: [y], z: [z],
	u: [u], v: [v], w: [w]
	});
	})
	.then(function() {
	_assertAxisRanges('after adding one cone outside range but with norm-0',
	[-0.72, 6.72], [-0.72, 6.72], [-0.72, 6.72]
	);

3D cone traces #2641

3D cone traces #2641

Conversation

etpinard commented May 16, 2018 • edited Loading

Things we need to decide on

Things we need to fix/add

Things that work ok, but could improve

Things that could wait future iterations

etpinard commented May 16, 2018

alexcjohnson commented May 16, 2018

alexcjohnson commented May 16, 2018

etpinard commented May 16, 2018

alexcjohnson commented May 16, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

kig May 17, 2018 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alexcjohnson commented May 16, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

etpinard commented May 17, 2018

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

alexcjohnson commented May 18, 2018

etpinard commented May 18, 2018

etpinard commented May 16, 2018 •

edited

Loading

kig May 17, 2018 •

edited

Loading